Re: Can anyone get Edge Copilot to work without logging in after July 2026?

On Tue, 7/28/2026 1:46 PM, Maria Sophia wrote:

> That means my machine would likely only work with LM Studio or Ollama.
> If I were to choose a setup for my old, 2009 Windows 10 box, it may be
>  a. LM Studio GUI
>  b. Phi 4 mini (3.8B) models
>  c. Which should work with ~2-3 GB disk & 16 GB RAM 
> 
> For now, I don't have the time to research how to install a local LLM.
> But if others are ahead of us, it would be good to know what they use.
> 

It's simple. Install LMStudio.

Then, look for a model in the list of open weight models it has.
Go and research the options it offers, for the one you think
will work in your setup. You could start with a Gemma 2.4B or 4B
for example, which should run on just about anything.

Inference can be done purely on the CPU, and it doesn't even
necessarily make good usage of the CPU. I'm running mine on
the CPU, as my GPU is not good enough for most all of them.

The model description, if it is a mixture of experts,
will tell you how big one of the mixture elements is.
The download files could be 58GB, while the MOE might
take 20GB of VRAM. And that would tell you what kind
of video card could be used. For example, the biggest Kimi
model, is 100GB+ for VRAM, so an RTX6000 at 96GB would not
be enough. It would take two RTX6000, and you have to
go off and research whether your video card choice supports
"sharding". Sharding does not apparently always work well.
And depending on the video card brand, may not work at all.
If a mixture of experts had different sizes models, it might
manage to run one model per video card.

It takes maybe a minute to load a large model, and it is
not purely I/O constrained. Initially it works at
the I/O rate, but once it gets to about 80% of the model
loaded, things slow down for some reason.

You don't need to do much, to get a text query window running.
It will be a lot harder, to research what exactly it is doing.
At the lower level, there is matrix math. At a higher level,
there is a strategy planner, and an ordered list of Experts
to run. And a few models, do not use a strategy planner, and
they bind an Expert to each token. The output rate on that one,
is well below 1 token per second, the approach then being a
"novelty" rather than being practical.

   Paul